Meta launches Muse Wonder models for long-running AI agents

On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full, two models designed for agents that can keep working on a task instead of stopping after one chat response. Meta says Wonder is available through the Meta Model API and Muse Code, while Wonder Full is available to Muse Code subscribers. The company plans to use Wonder in Muse, its own agent product, as access expands.

The most important distinction is how the models handle ongoing work. Meta describes Active Memory, which carries relevant details between conversations, and Active Focus, which lets an agent continue, try another approach after an error and ask for clarification or confirmation before an important action. The examples span insurance claims, leases, finance, law, health and software development. In practice, that could make an agent more useful for a multi-step brief, but it also makes permissions, audit trails and recovery behavior more important.

Meta’s model page lists a one-million-token context window for Muse Wonder and prices of $4 per million input tokens, $0.20 for cached input and $20 for output. Wonder Full is a subscription product rather than a generally priced API model. Those figures describe Meta’s current offering; they do not by themselves establish how the models compare with competitors in independent testing. TechCrunch reported that Meta supplied a model card, while Reuters placed the launch in the wider contest with Anthropic and OpenAI.

The announcement matters because the next competitive line is moving from answering questions to managing work over time. Memory and persistence can reduce repeated prompting and let an agent recover from a failed step, but they can also preserve a wrong assumption or expose information across tasks. Meta says the models are designed to avoid sharing private information without permission and to refuse harmful actions; a fuller safety report is still to come.

For developers, Muse Wonder is a concrete new option for agent experiments, with access limits and subscription conditions to check before adoption. For users, the useful question is not only whether an agent sounds capable, but whether it remembers the right things, knows when to stop and makes its actions visible.